Reciprocal Recommendation for Two-Sided Markets

170 concepts83 papers1 proposal

In a traditional recommendation system, like Netflix or Amazon, the relationship is a "one-way street." The system predicts whether a user will like a movie or a toaster, but the movie does not need to like the user back. However, in two-sided markets such as online dating, job recruitment, or peer-to-peer lending, the "item" being recommended is actually another active agent with their own set of preferences.

As Xia et al. describe, these environments require a "double yes": a successful match only occurs when both parties express mutual interest. This shift from unilateral relevance to bilateral consent transforms recommendation from a simple ranking problem into a complex coordination challenge.

The Geometry of Mutual Interest

In reciprocal systems, the goal is to predict the likelihood of a successful interaction, which is the product of two distinct directional preferences. Xia et al. propose calculating a final match score using the harmonic mean of two compatibility scores. This ensures that a pair only ranks highly if there is a strong likelihood of interest from both sides; a high score from one party cannot compensate for a zero from the other.

This reciprocal relevance is often modeled as a two-stage process. As Su et al. and Ribeiro et al. frame it, one side (the proactive party) initiates contact, and the other side (the reactive party) decides whether to respond. Consequently, the system must optimize for the joint probability of both the "send" and the "reply" actions. Guo et al. and the authors at LinkedIn note that this requires modeling not just what a searcher wants, but the historical engagement patterns and preferences of the candidate to ensure they are likely to respond.

Managing Capacity and Bottlenecks

A unique failure mode in reciprocal recommendation is the "superstar" effect. If a system recommends the most attractive candidates to everyone, those individuals become overwhelmed and cannot respond to every request. Nakada et al. argue that this creates bottlenecks that lower the total number of successful matches across the platform.

To solve this, researchers incorporate "matching theory" from economics. Tomita et al. suggest accounting for user "capacity"—the finite amount of time or attention a user has to engage with matches. By distributing recommendations more broadly across the ecosystem, systems can move toward a stable matching where no two parties would both prefer an alternative pairing, thereby improving overall market health.

From Individual Utility to Social Welfare

Because reciprocal recommendation involves multiple stakeholders, the optimization objective shifts from maximizing a single user's satisfaction to maximizing "social welfare"—the total number of successful matches in the system. Michael Jordan argues that this requires "intelligent markets" that blend computer science with microeconomics to solve scarcity and coordination problems.

This market-wide perspective introduces new challenges in fairness and efficiency:

  • Fairness: Jain et al. explore how to maximize Nash Social Welfare when both sides have conflicting valuations, a problem that is significantly more complex than one-sided resource allocation.
  • Redundancy: Yang et al. point out that showing both parties each other simultaneously can be a waste of limited recommendation slots. If one-sided exposure is enough to trigger a match, the other slot could be used to facilitate a different connection.

Dual-Perspective Modeling

Technically, these systems often employ symmetric architectures to capture the dual nature of the interaction. Kou et al. use dual-stream variational autoencoders to model user preferences and item exposure patterns in parallel. Similarly, Fu et al. utilize a dual-sided modulation mechanism where user interests filter item attributes while item "attractions" simultaneously modulate user profiles.

By treating the recommendation as a bilateral intervention rather than a unilateral suggestion, these models account for the fact that participants often have divergent motivations. As Lai et al. observe, a job seeker might prioritize salary and location, while an employer prioritizes specific technical skills. Only by modeling these distinct decision-making lenses can a system successfully facilitate a "double yes."

Key Foundational Work

Symmetric Collaborative Filtering Using the Noisy Sensor Model (2001) — contributed 2 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 89 papers
  2. Version 1 · 2026-10-04 · 93 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_reciprocal_recommendation_0a49e30c,
  title = {reciprocal recommendation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reciprocal-recommendation/rdr_0a49e30ceaf4480f8a1afe0e35c68e5c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_reciprocal_recommendation_0a49e30c_v2,
  title = {reciprocal recommendation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reciprocal-recommendation/rdr_0a49e30ceaf4480f8a1afe0e35c68e5c/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake